The warring states linguistic benchmark for hierarchical multimodal reasoning in digital heritage

Warring States Chinese, a crucial transitional stage featuring highly variable orthography, unconventional syntax, and fragmented discourse across heterogeneous media, reveals notable limitations in current large multimodal models. To address this, the Warring States Linguistic Benchmark (WS-LB) is proposed as a four-level hierarchical framework assessing orthographic, syntactic, semantic, and discourse understanding. Unlike prior benchmarks that evaluate isolated tasks, WS-LB uniquely models the cascading reasoning pipeline from visual perception to macro-level discourse. By capturing the dependency propagation of errors across these levels, WS-LB provides a highly realistic assessment of model capabilities in ancient manuscript analysis. Integrating cross-carrier visuals and annotations, WS-LB captures the cognitive demands of processing early Chinese manuscripts. Evaluated models (e.g., GPT-4o, Qwen2.5-VL, DeepSeek-V3) show strong surface perception but persistent deficits in semantic grounding and discourse reasoning. As the first comprehensive benchmark for this era, WS-LB establishes a computational foundation for multimodal ancient text processing and digital heritage preservation.